Applying Transfer Learning to QSAR Regression Models
Aiming at avoiding high costs in the production and analysis of new drug candidates, databases containing molecular information have been generated, and thus, computational models can be constructed from these data. The quantitative study of structure-activity relationships (QSAR) involves building predictive models that relate chemical descriptors for a compound set and biological activity with respect to one or more targets in the human body. Datasets manipulated by researchers in QSAR analyses are generally characterized by a small number of instances, which can affect the accuracy of the resulting models. In this context, transfer learning techniques that take information from other QSAR models to the same biological target would be desirable, reducing efforts and costs for evaluating new chemical compounds. This article presents a novel transfer learning method that can be applied to build QSAR regression models by Support Vectors Regression (SVR). The SVR-Adapted method for Transfer Learning (ATL) was compared with standard SVR method regarding values of mean squared error. From experimental studies, the performance of both methods was evaluated for different proportions of the original training set. The obtained results show that transfer learning is capable to exploit knowledge from models built from other datasets, which is effective primarily for small target training datasets.
KeywordsTransfer learning Support-vector regression Cheminformatics QSAR models Medicinal chemistry
The authors thank to the Brazilian funding agencies CAPES and FAPESP, and to IBM for financial support.
- 1.J. Devillers, A.T. Balaban, Topological Indices and Related Descriptors in QSAR and QSPAR (CRC Press, Boca Raton, 2000)Google Scholar
- 5.S. Wang, Z. Li, A new transfer learning boosting approach based on distribution measure with an application on facial expression recognition, in 2014 International Joint Conference on Neural Networks (IJCNN) (IEEE, 2014), pp. 432–439Google Scholar
- 7.L. Rosenbaum, A. Dörr, M.R. Bauer, F.M. Boeckler, A. Zell, Inferring multi-target QSAR models with taxonomy-based multi-task learning. J. Cheminformatics 5(1–2), 33 (2013)Google Scholar
- 8.T. Girschick, U. Rückert, S. Kramer, Adapted transfer of distance measures for quantitative structure-activity relationships and data-driven selection of source datasets. Comput. J. bxs092 56(3), 274–288 (2013)Google Scholar
- 10.C.A. de Moraes Lima, Comitê de máquinas: uma abordagem unificada empregando máquinas de vetores-suporte, Ph.D. Dissertation, Universidade Estadual de Campinas, 2004Google Scholar
- 11.T. Evgeniou, M. Pontil, Regularized multi-task learning, in Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (ACM, 2004), pp. 109–117Google Scholar
- 12.J. Garcke, T. Vanck, Importance weighted inductive transfer learning for regression, in Joint European Conference on Machine Learning and Knowledge Discovery in Databases (Springer, Berlin, 2014), pp. 466–481Google Scholar
- 13.S.C. Araujo, V.G. Maltarollo, D.C. Silva, J.C. Gertrudes, K.M. Honorio, ALK-5 inhibition: a molecular interpretation of the main physicochemical properties related to bioactive ligands. J. Braz. Chem. Soc. 26(9), 1936–1946 (2015)Google Scholar
- 14.M.O. Almeida, G.H. Trossini, V.G. Maltarollo, D.d.C. Silva, K.M. Honorio, In silico studies on the interaction between bioactive ligands and ALK5, a biological target related to the cancer treatment. J. Biomol. Struct. Dyn. 34(9), 2045–2053 (2016)Google Scholar